SKI Framework¶

Sovereign Knowledge Intelligence — an open neuro-symbolic architecture for AI compliance in regulated industries.

Status — spec v3.0; implementation on the v3.1.0-alpha line

A KG-grounded sovereign LLM is the primary reasoner on every verdict. The Symbolic Evaluator is repositioned as an independent verifier of the LLM's formalizable assertions. The Knowledge Graph is a typed semantic substrate with jurisdictional scope and effective-date intervals. The audit trail moves from deterministic replay to verifiable provenance. The architectural rationale is in RFC 0002 (Accepted; implemented in v3.0.0). All packages are on PyPI; see the roadmap for what's next.

Why SKI exists¶

Regulated industries — energy, finance, manufacturing, defense — need AI in core operational systems but cannot adopt frontier-model chatbots or cloud-hosted compliance APIs. The reasons are non-negotiable: operational data cannot leave the deployment perimeter, every decision must trace to a specific regulation, the audit story must survive cross-examination, and human judgment must remain the final authority. A rule engine satisfies the audit requirement but cannot reason about regulatory language. A frontier LLM can reason but ships data and cannot prove how it decided. SKI is the architecture that meets all four requirements simultaneously.

Pillar	What it means	How SKI delivers it
Sovereign	All evaluation runs on customer infrastructure; no data egress during inference	Local LLM runtime (Ollama, vLLM, or llama.cpp); model weights, KG, and ledger stay on the host
Knowledge	Regulations are a typed semantic substrate the system reasons over, not free text	Knowledge Graph with typed obligations, jurisdictional scope, exemptions, precedent, citations
Intelligence	An LLM that understands regulatory language, with its reasoning made auditable	KG-grounded local LLM (v3 primary); symbolic verifier on the formalizable subset; signed transcripts
Human primacy	AI supports human judgment, never replaces it	Five-verdict taxonomy with explicit `DISCRETIONARY`; high-tier rules require attestation tokens

What you get¶

:material-brain:{ .lg .middle } KG-grounded local LLM

Sovereign local model (Ollama / vLLM / llama.cpp), temperature=0, structured generation. Reads the typed KG slice for each rule; emits a verdict, reasoning, KG citations, and formalizable assertions.

:octicons-arrow-right-24: RFC 0002
:material-shield-check:{ .lg .middle } Symbolic Verifier

Independent cross-check of the LLM's formalizable assertions — numeric bounds, set membership, temporal windows. Disagreement is a first-class signal recorded in the ledger.

:octicons-arrow-right-24: Architecture
:material-file-document-multiple:{ .lg .middle } Knowledge Graph

Typed semantic substrate: obligations, definitions, exemptions, jurisdictional scope, effective-date intervals, precedent edges. Ed25519-signed; the runtime refuses to load an unsigned KG.

:octicons-arrow-right-24: KG schema
:material-database-lock:{ .lg .middle } Verifiable audit ledger

Postgres triggers reject UPDATE, DELETE, TRUNCATE. Each entry's hash chains to the prior; v3 adds signed LLM transcript, model weight hash, KG version hash, KG citations, verifier result.

:octicons-arrow-right-24: Replay docs
:material-shield-account:{ .lg .middle } Conformance suite

Black-box Provenance / Durability / Sovereignty tests citing the spec section they validate. Verifiable provenance is the audit contract every conformant runtime must produce.

:octicons-arrow-right-24: Conformance
:material-gavel:{ .lg .middle } Governance and RFCs

Lazy-consensus model with named maintainer teams. Architectural changes go through RFCs; the v3 pivot is RFC 0002.

:octicons-arrow-right-24: Governance

How SKI differs¶

A rule engine is fast and easy to audit but cannot reason about an actual regulation's language. A frontier-model chatbot can reason but ships data to a vendor and cannot prove how a decision was made. SKI sits between: an LLM reasons over a curated, signed knowledge graph that captures the regulation's structure; a symbolic verifier catches the subset the LLM can hallucinate on; every step is signed and chained into an append-only ledger. Each pillar is non-negotiable. Remove sovereignty and you cannot deploy in a regulated environment. Remove the knowledge graph and the LLM has no ground truth. Remove the symbolic verifier and an LLM can hallucinate a CLEAR verdict on a value that is plainly over the limit.

Quick start¶

Clone the repo, run scripts/setup.sh to generate TLS certificates and a .env file, start the Ollama container, pull the default open-weights model (qwen2.5:7b-instruct), bring up the rest of the stack with docker compose, send a sample telemetry record from examples/, and verify the audit ledger. The full walkthrough is in Getting started; the 10-minute newcomer path is in Your first rule.

Who's behind this¶

KpiFinity Inc. — a Calgary-based technology and consulting firm specialising in AI governance and compliance automation for regulated industries.

The specification is permissively licensed (CC BY 4.0) and is open to community evolution. The reference implementation and tools in this repository are Apache 2.0. The Knowledge Graph libraries for energy, finance, manufacturing, and defense are proprietary and licensed separately by KpiFinity.